Box-driven coarse-grained segmentation for stroke rehabilitation scenarios

Abstract

For complex stroke rehabilitation scenarios, visual algorithms, such as motion recognition or video understanding, find it challenging to focus on patient areas with slow motion amplitude and pay more attention to targets with drastic changes in light flow. Therefore, it can provide critical perspectives and adequate information for the above visual tasks using a semantic segmentation algorithm to capture the patient's area from the captured image. Currently, the weakly supervised segmentation algorithm based on bounding boxes tends to utilize existing image classification methods. They can perform secondary processing on the internal images of boxes to obtain larger areas of pseudo-label information. In order to avoid the redundancy caused by algorithm concatenation, this paper proposes an end-to-end weakly supervised segmentation algorithm. In this method, a U-shaped residual module with variable depth is designed to capture the deep semantic features of images, and its output is integrated into the target matrix of the NCut problem in the form of blocks. Then, the region of the target is indicated by solving the sub-minimum eigenvector of the generalized eigensystem, and the segmentation is realized. We conducted experiments on the PASCAL VOC 2012 dataset, and the proposed method achieved 67.7% mIoU. On the private dataset, we compared the proposed method with similar algorithms, which can segment the target area more intensively